Cowork

Sources#

• Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next
• How Anthropic's product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
• The Founder's Playbook: Building an AI-Native Startup

Summary#

Anthropic's knowledge-work agent product, sibling to Claude Code. Where Claude Code targets work whose output is code, Cowork targets work whose output is non-code: slide decks, inbox triage, customer dossiers, launch docs, meeting prep. Co-led on product by Cat Wu.

Positioning (Cat Wu's framing)#

• Claude Code (CLI) — one-off coding tasks; CLI gets new features first; "the most powerful of all the tools"
• Claude desktop / Claude Code in IDE — front-end work where you want a preview pane; better for non-technical users uncomfortable in terminal
• Claude Code mobile / web — kicking off tasks AFK; "missing product" before it shipped — people would tether laptops to phones outdoors
• Cowork — anything where the output isn't code

How Cat Wu uses it#

Connects Google Calendar, Slack, Gmail, Google Drive — "the more context, the better." Example workflow: feeds a draft outline + relevant Twitter / launch-room links into Cowork at night, returns to a 20-page polished slide deck the next morning, iterates with one round of feedback. The Anthropic design system is loaded so the deck looks like a designer made it.

Why people use it (per Cat Wu)#

The "Applied AI" team — second-largest internal token spender after engineering — uses Cowork for:

• Customer-meeting prep ("dossier" workflow: yesterday's questions, action items, ETA research)
• Inbox-zero / Slack-zero triage
• Launch plans, internal docs, customer-tailored decks pulled from Salesforce + Gong + Slack

Where it ties into the larger Anthropic roadmap#

• Boris Cherny cites Cowork as the venue for general-purpose MCP and computer use integration — same MCP connectors as Claude AI, plus Anthropic's lead on computer use for software lacking MCPs ("quite good" on Opus 4.7).
• "We have no more manually written code anywhere at the company" — Cowork covers the non-code half of that claim (SQL, decks, ops docs all model-generated).

In the founder playbook (May 2026)#

The Founder's Playbook positions Cowork as the operational layer for AI-native startups across every stage:

• Idea stage — automate customer outreach + scheduling. Cowork uses the validated target profile to research prospects, draft personalized outreach, connect to Gmail/Calendar via MCP, manage threads, run day-7 follow-ups.
• MVP stage — feedback-loop logistics: outreach to early users, scheduling feedback sessions, structured bug-and-feature intake, weekly synthesis.
• Launch stage — replaces founder attention at the operational layer: audit founder's recurring tasks, categorize automate/delegate/founder-only, build workflow logic for automation candidates. PM operating system: sprint scheduling, bug-triage routing, weekly metrics compilation.
• Scale stage — enterprise-grade operations: ticket routing, escalation workflows, documentation triggered by product changes, renewal tracking, reporting cadences. Plus GTM tactical execution: content pipelines, outbound sequences, analyst briefing logistics, CRM hygiene.

The playbook's framing: Cowork + Claude Code together give "a small team the support posture of a much larger organization." Skills are highlighted as the codification surface for recurring founder workflows.

The counterfactual: OpenAI didn't split (July 2026)#

Cowork's existence encodes a bet — that the natural seam between a coding agent and a knowledge-work agent is what the output is. OpenAI faced the identical problem (Codex adoption inflecting among non-developers) and made the opposite call: one shared harness, with only git-state visibility, diff-forward display, and sandboxing defaults differing between the Codex and ChatGPT Work surfaces. Akshay Nathan's stated reason is that role boundaries are dissolving, so "trying to draw a hard boundary based on who you are is gonna be tough… we shouldn't box them in."

Note the seams aren't identical: Cat Wu's rule splits on output type, which dodges part of Nathan's objection — but it still asks the user to pick a surface before they know what the task will produce. Note also the dates: this page's material is May 2026 and Nathan's is July 2026, so some of the gap may be lag rather than disagreement. Neither side offers evidence; treat as two live product bets. Full comparison at Shared Harness, Differentiated Surfaces.

Key quotes#

• (Cat Wu) "If I'm building something where the output is code, I'll use Claude Code or desktop or Claude Code on mobile. And if the output is anything that's not code, I'll use Cowork for it."
• (Boris Cherny) "Claude design I think is a really good example. It's pretty good today. It's going to get a lot [better]." (Referring to Cowork-class products in his vision answer.)

Connections#

• Claude Code — sibling product
• Shared Harness, Differentiated Surfaces — the OpenAI counterfactual: one harness with per-surface UX opinions instead of a product split by output type
• Claude Design — sibling Anthropic Labs product; Cowork's output is decks/docs, Claude Design's is visual design/prototypes; Boris Cherny flagged Claude Design as a vision example here
• Cat Wu — Head of Product
• Boris Cherny — names it as an area of investment
• Anthropic — vendor
• Engineer PM Convergence — Cowork plus Claude Code is what makes "everyone codes" + "everyone runs ops via agents" possible across non-engineering functions
• AI Employee Framing — Cowork's deployment surface (Gmail, Slack, Calendar, Drive) is where "AI as employee" framing decisions get made; HBR research suggests how these get framed changes accountability and review quality
• Human-AI Accountability Redesign — non-code agent output (decks, dossiers, inbox triage) lacks compiler/test verification, so accountability redesign matters more, not less, than for Claude Code
• Agentic Misalignment (AM) — Cowork-style deployments (long-context, tool-using, weak per-action oversight) are exactly the AM eval setting
• AI-Native Startup Lifecycle — Cowork is the operational layer across all four founder stages
• Founder as Agent Orchestrator — Cowork plus MCP integrations is what makes the orchestrator role tractable for a solo founder
• Compounding Data Moat — Cowork runs the enterprise-support layer that compounds workflow lock-in at Scale
• Conversation Artifacts — Cowork's outputs (decks, dossiers, inbox triage) are among the artifacts the AEI Cadences report classifies; Cowork data is folded into "Claude conversations" throughout
• Problem-Solution Fit Discipline — Cowork's Idea-stage role is research, interview-framework audit, and post-interview synthesis — the operational layer of the validation discipline
• Fiona Fung — leads engineering + product for Cowork (alongside Claude Code); onboarded small-business customers onto Cowork to feel onboarding pain (Dogfooding as Product Discipline)

Open Questions#

• What's the eval discipline for Cowork-class outputs? Cat Wu says memory benefits a lot from evals; unclear how slide-deck quality is measured.

Resolved Questions#

• How does Cowork's harness compare to Claude Code's? Both surface skills, MCP, sub-agents — but the failure modes for non-code output differ (no test suite, no compiler, no diff to review). Answered: Verifying Without a Compiler: Cowork's Harness vs Claude Code's, and Why the Slice Verifier Stays — same primitives, opposite verifier rungs, so the harness weight redistributes: Claude Code leans on a post-hoc deterministic verifier stack that both catches errors and bounds damage pre-merge; Cowork substitutes judgment-encodings (the loaded design system as the nearest thing to a style linter, evals/LLM-judges, human review at decision checkpoints) and makes the pre-action classifier gate load-bearing, because errors ship directly into live SaaS state with no red test in between. Failure modes split loud (build breaks) vs silent (a polished deck that reads fine — the failures-that-look-like-success class), which is why accountability redesign matters more here, not less.

Derived#

• Verifying Without a Compiler: Cowork's Harness vs Claude Code's, and Why the Slice Verifier Stays — the harness comparison with Claude Code: what replaces the compiler on the non-code surface

Sources#

• How Anthropic's product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
• Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next
• The Founder's Playbook: Building an AI-Native Startup